Skip to content

[BLS, VLLM and Python Backend] #8917

Description

@protonicage

Description
I have problems when using BLS to communicate with the VLLM backend. The python BLS model (which calls a VLLM model)
itself is async but decoupled=False and also make an async_exec(decoupled=True) call to the VLLM model.

This does not work for reasons I can not comprehend completely, but it has to do with the asyncio backend which gives me an "invalid argument" error when i try to await the requests.

I found this hint in the python backend:
https://docs.nvidia.com/deeplearning/triton-inference-server/user-guide/docs/user_guide/bls.html#limitation

For me this mean calling an decoupled python model with async_exec will not work, but this would mean your default implementation of vllm would not be usable with BLS calls.

But then I found also this:
https://github.com/triton-inference-server/python_backend#decoupled-mode

Is this limitation of BLS still valid?

Triton Information
26.06 NGC Container

Are you using the Triton container or did you build it yourself?

I use a NGC container.

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions